LEVERAGING MACHINE LEARNING

WITHIN ANTI-MONEY LAUNDERING

TRANSACTION MONITORING
Artificial Intelligence (AI) and one
of its key components, Machine
Learning (ML) are innovative and
disruptive technologies that are
widely used within many industrial
sectors, including life sciences,
automotive, aerospace and defense.

However, despite numerous potential Nevertheless, when coupled with hybrid

applications within the financial services models involving human resources supported
sector, specifically within Anti-Money by AI, AML solutions can be developed
Laundering (AML), adoption of AI and ML that provide significant benefits, including
has been relatively slow. This has been, a reduction in compliance costs and
in part, due to the following: (1) limited improvements in the transaction monitoring
comprehension of the application of AI process leading to greater overall efficiency.
and ML within AML compliance programs;
(2) the notion of ML being a black box
where the inner workings are not clearly
understood by regulators and compliance
officers; and (3) the cautious approach
taken by regulators requiring compliance
officers to understand and validate how
outcomes from AML models are arrived at.

2 EVOLVING AML JOURNEY

ML APPLICATIONS FOR AML
TRANSACTION MONITORING
AI is a field of computer science focused on the development
of computers to carry out activities typically executed by
humans, specifically activities requiring humans to take
decisions and rely on their intelligence.
There are several component technologies
within AI all with numerous applications.
For example, Natural Language Processing
(NLP), through which computers process
human language, is used for translating
payments in payment screening, processing
mortgage applications and negative news
screening (see Figure 1).
ML is a sub-field of AI where computers
are able to learn when exposed to new
data without being explicitly programed.
The potential applications for this within
suspicious activity monitoring and
Transaction Monitoring (TM) are vast.
A common TM challenge is the generation
of a vast number of alerts requiring costly
operations teams to triage the alerts.
With increasing customer numbers, and
increasing numbers of transactions, TM
alert volumes have increased significantly
at financial institutions (FIs). This increase
in alerts has resulted in an upsurge in the
number of personnel required to triage and
process them. However, in leveraging ML at
the alert triage stage, alerts can potentially
be suppressed, hibernated (set aside for
later examination) or even, at some time
in the future, closed automatically.
ML can teach computers to detect and
recognize suspicious behavior and classify
alerts following a risk-based approach
as being of high, medium or lower risk.
Applying rules to these alert classifications
can facilitate hibernation and auto closure
of alerts, allowing human resources to
supervise machines that triage these alerts
rather than reviewing alerts manually.
This change is illustrated in Figure 2.
Machines can be taught to recognize,
score, triage, enrich, close or hibernate
alerts. This can help focus experienced
human resources on complex problems
while computers perform simpler tasks
for them. By using AI across the entire TM
process, alerts can be auto-triaged leveraging
ML; they can be auto-enriched with account,
transaction, and customer information from
external and internal data sources leveraging
intelligent automation and NLP; and narratives
for Suspicious Activity Reports (SARs) can
be auto-generated using Natural Language
Generation (NLG). Institutions leveraging AI in
this way can reduce dependencies on human
operators to perform routine tasks, reduce
the total time it takes to triage alerts, and
allow personnel to focus on more valuable
and complex activities.
EVOLVING AML JOURNEY 3
 Figure 1. Illustration of some AI Applications and Components
Other components include, Genetic Algorithms and Swarm Intelligence. This document largely
focuses on ML for AML TM.

DEEP EVIDENCE
Input Hidden Output LEARNING BASED Documents Possible Answers
Layer Layers Layer

RECOMMEN- ML
DATION SYSTEMS Linear Clustering
Data Recommendation ENGINES Regression
Models

PREDICTIVE PRESCRIPTIVE
Available Predictions ANALYTCS AI ANALYTICS Available Assessment
Features Features

Advice

NLP/
NLG
TEXT MINING

Document Data Data Language

Source: The Artificial Intelligence Ecosystem, Narrative Science

Figure 2. Introducing ML into AML TM Alert Triage

With the introduction of ML into AML TM alert triage, SAR conversion rates should improve from the
current unacceptable rate of ~1% in the banking sector.

DETECT RECOGNIZE PROCESS

ROTATING TO THE NEW

Escalated Alerts
SARS
Automation 1%
TXNS
Investigations
Data from
Multiple Automation Auto-enriched Alerts
Sources TM Analysts Closed Alerts

Auto-triaged
Alerts Automation

Auto-closed / Hibernated Alerts

High Potential to Reduce Human Effort

Source: Accenture, September 2017

4 EVOLVING AML JOURNEY

CHALLENGES TO
IMPLEMENTING ML
Through Accentures work with global FIs we have identified
a number of implementation challenges for financial services
firms seeking to implement ML as part of AML TM solutions.

Figure 3. The Challenges to Implementing ML as part of a TM Solution

DQM
and Profile
Refresh

Lack of Lack of a
Straightforward 360-degree
Processes View of the
Customer
CHALLENGES
FOR ML & AI

Limited
Limited
Knowledge of both
Regulatory
Financial Services
Appetite
and ML

Source: Accenture, September 2017

EVOLVING AML JOURNEY 5

These challenges include poor data quality,
low Know Your Customer (KYC) profile
refresh rates, lack of a holistic view of the
customer, limited knowledge of both financial
services and ML, regulatory expectations
and appetite, and lack of straightforward
processes to follow for implementation.
1. Data quality management and profile
refresh. Data Quality Management (DQM)
is a required activity for all monitoring
systems and analytics. It is not unique in
any way to ML applications. Both static
files, for example KYC files, and dynamic
data held by FIs on their customers
transactions have seen low completeness
ratios in areas such as missing payment
information and high error rates in recent
years. Additionally, there is also a general
lack of data traceability and data lineage.
Some FIs address data issues on an
alert and/or profile basis when alerts
aregenerated as false positives or noise.
This method reduces reoccurrence
of noise or false positives. However,
others prefer large-scale, one-off
data reconciliation/refresh exercises.
During the last few years, many FIs
have undertaken large and costly data
remediation projects to improve data
and have implemented frameworks to
manage data quality. However, data
quality remains an ongoing problem.
Tothe challenge of missing and outdated
data, there is an opportunity to tie profile
refreshes into sales and marketing
exercises. Relationship managers
can use profile refresh opportunities
to reach out to their customers and
identify cross-selling opportunities.
6 EVOLVING AML JOURNEY
2. Lack of a 360-degree view of the
customer. Another challenge in
implementing ML-supported TM programs
is the inability for FIs to build a 360-degree
view of the customer. For example,
financial services firms do not have the
global freedom to share information about
their customers to build a comprehensive
network. Furthermore, FIs do not formally
collaborate on AML as this is a cost.
Building a 360-degree view would require
collaboration across FIs and information
agencies that does not exist today.
Currently, once FIs decide to file an SAR
and submit it to the regulator, there
is limited feedback and intelligence
returned. In addition to sharing
information securely between FIs and
regulators, external (to KYC) data sets
can be used in areas such as marketing
data, which has rich datasets on customer
behavior not generally used for AML
behavior modeling and alert enrichment
and/or triage. However, there have been
some developments in recent years
that are paving the way to alleviate this
challenge of data sharing. Regulators are
increasingly leaning toward data sharing
between banks. Relevant regulations
encouraging data sharing are US Patriot
Act 314 a and 314 b and PSD2. The UK
Treasury is also encouraging sharing
of data via an Open Banking API/Open
Banking Working Group. However, data
ownership, medium (e.g. cloud) and the
extent of this data sharing needs to be
defined. It is a matter of time before large
amounts of transactional data will become
available on intrabank data clouds.
3. Limited knowledge of both financial 5. Lack of straightforward processes.
services and ML. Another reason why ML is a relatively new technology and
ML has not been as widely applied there are few established, straightforward
within the financial services industry is processes to follow to implement it.
that very few people are experts in both Without knowing what to look for, teaching
ML techniques and financial services. systems to detect certain types of financial
As a result, there have been fewer crime can be tricky. For example, how
applications targeting financial services does one teach a system to recognize
problems from start-ups and established terrorist financing? There is a carousel
vendors, limiting acceptance of ML within process for fraud but nothing similar for
the financial services sector (especially terrorist financing (nothing, that is, other
for financial crime prevention). than name matching against terrorist
lists). While some of these problems are
4. Limited regulatory appetite. FIs have
better suited to unsupervised learning,
held back from implementing ML solutions
model validators should be sure about
because regulators expect models that can
the desired outcomes.
be clearly understood. They typically insist
that all choices, limitations, and results
be documented before such models are
implemented. In the regulators view, the
models should be designed in a way that
allows for results to be reproduced given
the same input data. This is not always
possible with ML algorithms.

Additionally, there is some evidence that

large vendors are, in effect, acting as
leading regulators. This is because some
regulators recommend specific solutions
for TM that are ineffective and generally
need to be replaced by intelligent and
adaptive solutions. These large players
dominate the technology space and,
consequently, there is only a narrow field
of play for start-ups. Start-ups tend to
need support from consultancies and
other third parties that help banks with
their evaluations and implementations.

EVOLVING AML JOURNEY 7

MODEL VALIDATION AND
BLACK BOX TECHNIQUES
FOR ML AML MODELS
Any statistical model chosen by an FI for TM needs to follow steps
similar to those illustrated in Figure 4.
These steps are recommended by the Office
of the Comptroller of the Currency (OCC)
for model risk management.1 Extensive
documentation is required for models
selected and the regulators expect FIs
to be able to re-create results and explain
exactly how their models work.
The introduction of ML opens up numerous
new techniques for TM, all constructed with
mathematical equations. Some of these
algorithms or techniques are simpler to
understand than others. Simpler algorithms
include decision trees while more complex
algorithms may comprise neural networks.
Algorithms that are predictable, traceable,
and simple to explain are white box
techniques. On the other hand, more
complex algorithms are known as black
boxes, as the mathematical calculations
underlying these techniques are far too
dense and complex to be able to recreate
results and provide rationale for the learned
parameters (such as weighting). Figure 5
illustrates how model validation becomes
more complicated as techniques move
from white box to black box.
8 EVOLVING AML JOURNEY
Currently, there is no straightforward
method to validate complex or black box
ML model choices for internal auditing and
for external regulatory purposes. Black box
ML techniques can be tested by adopting
(and adapting) the Turing test, whereby
human analysts are trained with the same
samples as machines and an observer
evaluates results to determine precision
and recall. At the same time the observer
needs to determine which results were
produced by a human and which were
produced by an ML system. This process
could be undertaken as an interbank activity,
possibly with the involvement of regulators.
However, just as current TM systems
require their scenarios and segments to
be updated, ML-based TM systems will
also have to be updated periodically. This
can be more complicated than standard
TMsystem maintenance and ongoing testing
is required. Realistically, such testing and
updating should be consultancy or vendor
driven as part of an ongoing maintenance
contract. The slightest change in the
type of data that once produced desired
results can render the system unreliable.
Figure 4. Summary of Model Risk Management Guidance from the OCC

SENIOR MANAGEMENT OVERSIGHT

MODEL RISK MANAGEMENT

CONCEPT ONGOING OUTCOME

SOUNDNESS MONITORING ANALYSIS
GUIDANCE

All model choices Benchmarking (precision, Sensitivity analysis

Mathematical calculations recall, harmonic mean) Back testing (actual vs. forecast)
Key assumptions User feedback Testing approaches
Data, selection of inputs Limitations assessment Test data and outcomes
Limitations Time periods for testing

PRE-IMPLEMENTATION AND POST-IMPLEMENTATION

Source: Supervisory Guidance on Model Risk Management, Board of Governors of the Federal Reserve System,
the Office of the Comptroller of the Currency

Figure 5. Complexity Scale for Model Validation of ML Algorithms

SIMPLER MODEL VALIDATION FOR REGULATORS COMPLEX

Money Laundering?
+ Hidden
Input Output
Volume CRR

High High
_ Hyperplane
Classification and Artificial Neural
Regression Tree (CART) Support Vector Machines (SVM) Network (ANN)

WHITE BOX BLACK

Source: Accenture, September 2017

EVOLVING AML JOURNEY 9

HOW ACCENTURE CAN HELP
Accenture provides services to global FIs to implement both
ML and robotic process automation (RPA) programs.

We have extensive capabilities in this area,
including analytics and digital centres of
excellence and subject matter specialists to:
Identify and analyze business opportunities
Assess the impact of possible solutions
(such as updating procedure manuals
versus adopting ML)
Deliver innovative bespoke and industry
leading technology solutions, including
implementation of TM systems, AML
model validation, and operating model
transformation to improve the human
and virtual workforce
We have vendor alliances, both with
Fintech start-ups2 and with leading firms.
Accenture consultants meet on a regular
basis with new vendors, help vendors build
their capabilities in the financial crime
prevention space, and evaluate them
to deliver greater value to clients. Within
the financial crime space, Accenture
has established an effective and efficient
methodology to allow our clients to meet
regulatory expectations and reduce
volumes of TM and Name Screening
alerts. This methodology is illustrated
in Figure 6 below.

Figure 6. Summarized Steps for Rapid Benefits Realization from ML and RPA

BUILD AND TEST

IDENTIFY DATA SCOPE ROLL OUT,
DEFINE ADVANCED
SOURCES AND IMPLEMENTATION MAINTAIN AND
OPPORTUNITIES ANALYTICS
SYSTEMS TRANCHES UPDATE
APPLICATIONS

Identify pain Analyse available Outline immediate, Enhance and Periodically review
points and required data mid-term and fix data and tune
Identify processes Perform data long-term goals Algorithm and applications for
and procedures quality analysis Identify the "art of vendor selection ongoing suitablility
that can be (e.g. completeness the possible" for Build, deploy User training and
automated ratios) automation and ML and test proofs feedback to
Build a Develop delivery of concept improve usability
consolidated data approach and Model validation and
repository road map regulatory proving
Define underlying Implement and roll
data management out model
framework
Monitor business
benefits

Source: Accenture, September 2017

10 EVOLVING AML JOURNEY

Within the financial crime space, Accenture
has established an effective and efficient
methodology to allow our clients to meet
regulatory expectations and reduce
volumes of TM and name screening alerts.

EVOLVING AML JOURNEY 11

CONCLUSION
Many banks have started implementing business process
automation in the form of RPA and see AI and particularly
ML as the next step in the journey to greater efficiency and
effectiveness. Some banks, however, believe they should walk
before they run, meaning that robotics solutions should be
in place before AI ML solutions are considered.

In fact, robotics and AI ML can exist The appetite for automation is

independently of each other and each
can support the others capabilities.
Robotics can be used to train AI ML
models and AI ML models can be used
to add decision-making (in the form of
NLP) or reading (via optical character
recognition) capabilities to robotics models.
increasing rapidly, but challenges
remain. Implementing ML may improve
eventtoSAR conversion rates but
there will always be a need for human
involvement. Hybrid human/AI models
are likely to appear in the near future and
should represent the next generation
of AML TM alert operations teams. The
adoption of ML and, ultimately, AI in AML
is only the beginning of financial services
compliance and resilience automation.

12 EVOLVING AML JOURNEY

ABOUT THE AUTHORS
Samantha Regan
Samantha is a Managing Director, and is
the North America Lead for the Regulatory
Remediation & Compliance Transformation
group within Accentures Finance & Risk
practice. Based in New York, she has
17years of global experience working
with C-suite executives and their businesses
in compliance and regulatory initiatives.
Heather Adams
Heather is a Managing Director, Accenture
Finance & Risk. Based in London, she leads
Accentures UK & Ireland Resilience team,
developing capabilities and supporting
clients in their efforts to be resilient to
a range of external and internal threats,
including financial crime. Heather has
extensive experience in delivering large-
scale complex business change for banks
and has worked with senior leaders to
define and implement financial crime risk
strategies to drive high performance.
Philippe Guiral
Philippe is a Senior Manager, Accenture
Finance & Risk, based in New York. He leads
Accentures North America Financial Crime
Group, and has over 12 years of experience,
defining capabilities and strategies to
support clients in their financial crime
prevention programs in both the UK and
the US. Philippe specializes in delivering
large-scale complex business and
technology financial crime change
programs to drive high performance.
Saad Choudri
Saad Choudri is a Technology Consulting
Principal, Accenture Finance & Risk,
based in London. Saad has over 12 years
of international multisector experience
as a change professional with a special
focus on financial crime prevention and
artificial intelligence.
EVOLVING AML JOURNEY 13
14 EVOLVING AML JOURNEY
The appetite for automation is
increasing rapidly, but challenges remain.
Implementing ML may improve event-to-SAR
conversion rates but there will always
be a need for human involvement.

EVOLVING AML JOURNEY 15

REFERENCES
1. Supervisory Guidance on Model Risk
Management, Board of Governors of
the Federal Reserve System, the Office of the
Comptroller of the Currency, April 4, 2011.
Access at: https://www.occ.treas.gov/news-
issuances/bulletins/2011/bulletin-2011-12a.pdf
2. FinTech Innovation Lab, London:
http://www.fintechinnovationlablondon.co.uk/
ACKNOWLEDGMENTS
The authors would like to thank the following
Accenture employees for their contribution
to this document:
Garikai Mparutsa
Zhana Gali
STAY CONNECTED
Accenture Finance and Risk
www.accenture.com/financeandrisk
Finance and Risk Blog
financeandriskblog.accenture.com/
Connect With Us
www.linkedin.com/showcase/16183502/
Follow Us @AccentureFSRisk
www.twitter.com/AccentureFSRisk
Copyright 2017 Accenture
All rights reserved.
Accenture, its logo, and
High Performance Delivered
are trademarks of Accenture.
ABOUT ACCENTURE
Accenture is a leading global professional
services company, providing a broad range of
services and solutions in strategy, consulting,
digital, technology and operations. Combining
unmatched experience and specialized skills
across more than 40 industries and all business
functionsunderpinned by the worlds largest
delivery networkAccenture works at the
intersection of business and technology to
help clients improve their performance and
create sustainable value for their stakeholders.
With more than 411,000 people serving clients
in more than 120 countries, Accenture drives
innovation to improve the way the world works
and lives. Its home page is www.accenture.com
DISCLAIMER
This document is intended for general
informational purposes only and does not
take into account the readers specific
circumstances, and may not reflect the most
current developments. Accenture disclaims,
to the fullest extent permitted by applicable
law, any and all liability for the accuracy
and completeness of the information in this
document and for any acts or omissions
made based on such information. Accenture
does not provide legal, regulatory, audit,
or tax advice. Readers are responsible for
obtaining such advice from their own legal
counsel or other licensed professionals.
173270